How every verdict here is produced

We start from the token, not from a directory. A project qualifies if it has a live token with a market cap and a website it declared itself. Then we go and test that website.

The steps

1. Collect tokens that declare a website.
2. Open the site and read what it claims. Does it say MCP? Does it say x402?
3. Look for the machine surface: /.well-known/mcp.json, /.well-known/x402, /.well-known/agent-card.json, any MCP URL printed on the page, then the six common paths.
4. Speak real MCP to whatever we find: initialize, then tools/list.
5. Publish what came back.

What each verdict means

Own agent live: The project built a machine surface and it answered.
Claims, no answer: The project's own site says MCP or x402, and nothing answered at any endpoint we could find. This is the one that matters.
Docs vendor: Something answered, but it is a documentation host's built-in feature, not a product this project built. We separate these because counting them as agents would flatter everyone using GitBook.
No agent: Site is up, claims no machine surface, and none was found. Not a criticism. Most projects do not need one.
Auth wall / paywalled: Reachable, but wants credentials or payment first. Both are working as designed.
Site down: The website the token itself points at did not respond.
No own site: The token's declared website is a marketplace or social page, so there is nothing of its own to test.

How we handle market cap, because it lies

Two failures we hit and corrected rather than shipped.

Artifacts. One bad trading pair can report an absurd number. PYTH has 21 pairs: the median says $59.7M, one pair said $310.8B. We take the median across pairs, never a single pair, and every corrected figure says so on its page.

Ticker collisions. A token on Base calling itself SOL is consistently worth $16.9M across 16 pairs. No median catches that, because the number is real, the identity is not. We flag those by name.

We also show liquidity beside every market cap. When a market cap is hundreds of times its liquidity we say so, because that gap is the difference between a number and money you could actually leave with.

What we do not claim

Not finding an endpoint is not proof there isn't one. It can be at a path we did not guess, behind a login, or blocked to our range. It means one thing: at the timestamp on that page, from our network, nothing answered where the project pointed us.

We do not measure quality. Only whether a thing is there and what it says it can do.

Most projects here have no agent and that is fine. The finding is not that everyone is lying. It is the gap between the 34 that advertise a machine surface and the 5 that built one.

If we got yours wrong, tell us. We will re-run it and publish the correction on the same page.

Reproduce it

POST any endpoint with {"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"you","version":"1"}}} then tools/list. A real server returns serverInfo and a tool list. You do not need us.